Comparative evaluation of deep learning architectures for microbial colony classification in microbiological imaging

Area of Science:

• Microbiology
• Computer Science
• Bioinformatics

Background:

• Microbial colony visual interpretation is crucial in microbiology.
• Automated image analysis offers consistency and decision support in high-throughput screening.
• Limited comparative data exists for deep learning Convolutional Neural Networks (CNNs) in microbial identification.

Purpose of the Study:

• To conduct a head-to-head comparison of six CNN architectures for microbial colony classification.
• To evaluate the performance of different CNN models on a standardized dataset.
• To identify optimal deep learning models for automated microbiological image analysis.

Main Methods:

• Six CNN architectures (AlexNet, SqueezeNet 1.1, ResNet18, ShuffleNetV2, EfficientNetB0, MobileNetV2) were compared.
• The AGAR dataset, comprising 86,045 colony crops from five species, was utilized.
• Models were trained and validated using a unified protocol, with performance assessed by accuracy, precision, recall, and F1-score.

Main Results:

• MobileNetV2 achieved the highest accuracy (99.40%), followed by EfficientNetB0 (99.38%).
• ShuffleNetV2 (99.19%) and ResNet18 (99.16%) also demonstrated strong performance.
• AlexNet had the lowest accuracy (98.74%), indicating modern lightweight architectures performed best.

Conclusions:

• Modern CNNs, especially MobileNetV2 and EfficientNetB0, provide accurate microbial colony classification.
• These models show potential for improving automated microbiological image analysis.
• Integration into medical equipment can reduce manual labor, expedite diagnosis, and lower costs.